The Analysis of COVID-19 Surveillance Data: What can we Learn from Limited Information?
53 Pages Posted: 2 Jul 2020
Date Written: June 12, 2020
What can we learn about the transmission process as the COVID-19's pandemic unfolds when the only available information is daily counts of confirmed cases and tests? In this paper we propose to go old school and apply simple time series treatments to filter away seasonality and atypical values. Then, we use the trend and cycle components of the resulting time series to emulate the underlying infection process (the data-generating process) that gave rise to the noisy signal we observe in the data. We use these trends to compute the effective reproduction number Rt and the test positivity rate ρt, and propose a joint analysis of trajectories over (Rt,ρt) space to have a graphical assessment of the current status of the infection process. Before applying our method to country data, we test it using simulations. We find that although the level of epidemiological indicators is systematically biased when based only on surveillance data, our method allows us to reduce the root mean squared error and the probability of type II errors, particularly when testing strategies are poor. Moreover, the joint analysis of Rt and ρt manages to reduce the probability of type I and type II classification errors to below 0.5% in our simulations. Our country analysis, on the other hand, shows that daily counts of cases and tests exhibit strong seasonal and atypical components, two types of stochastic innovations that are orthogonal to the infection process and that, if left untreated, can produce spurious dynamics in the behaviour of Rt and ρt.
Note: Funding: We did not receive any funds for the research done in this paper. Declaration of Interest: The authors declare no conflict of interest.
Keywords: COVID-19, Statistical Simulation, Time Series Analysis, Agent-Based Modeling, Health
JEL Classification: C15, C22, C69, I10
Suggested Citation: Suggested Citation